The Mogao caves in Dunhuang, China, comprise 492 painted Buddhist cave temples dating from the 4th to 14th centuries, representing a vast resource for studying the history of art, architecture, religion, and cultural exchange along the Silk Road. Cave 465, unique in its Indo-Tibetan tantric Buddhist style, presents a challenge for dating due to its stylistic uniqueness. Historians debate its construction date, with suggestions ranging from the 9th to the 14th century. This study utilizes a novel interdisciplinary approach that combines advanced remote sensing techniques with machine learning and material analysis to address this dating debate. Distinguishing between the construction date of the caves and the date of the paintings is crucial, as repainting and restoration are common. Archaeological evidence, stylistic comparisons, and detailed analysis of painting materials can provide valuable insights. Recent advances in imaging spectroscopy allow for non-invasive, large-scale analysis of wall paintings, significantly increasing the amount of data available for analysis. This study tackles the challenge of analyzing this large dataset through the development and application of a machine learning-based approach for material identification, coupled with traditional art historical and archaeological analyses.

Scientific analysis has been applied to art historical studies since the 19th century, but traditional methods were limited to small, isolated areas. Advances in detector technology have enabled imaging spectroscopy, collecting millions of spectra simultaneously for non-invasive analysis of large areas. Clustering methods have been widely applied in various fields, including remote sensing, astronomy, and medicine, but automated clustering of large spectral image cubes from paintings remains under-explored. Existing methods often require manual intervention, leading to inexact classification. For example, the 'spectral angle mapper' is sensitive to spectral shape but not intensity, which can be a drawback when illumination is controlled. The use of unsupervised methods such as Kohonen Self-Organizing Map (SOM) has been proposed, but not for large-scale datasets or with high spectral resolution. This study addresses these limitations by developing a novel clustering algorithm that automatically processes large reflectance spectral imaging datasets, effectively clustering pixels of similar spectral reflectance (both shape and intensity) for material composition analysis.

The researchers used the PRISMS system, an in-house developed automated remote spectral imaging system, to capture large areas of wall paintings in Mogao Cave 465. PRISMS uses a Jenoptik CCD camera, a filter wheel with 10 filters (400-880 nm), and a telescope to collect high-resolution spectral images from a distance. The system automatically calibrates the images to generate spectral reflectance image cubes. The study developed a novel clustering algorithm using Kohonen Self-Organizing Maps (SOM) to analyze the large spectral imaging datasets. The algorithm consists of three main steps: 1) unsupervised clustering of an initial image cube to create a reference spectral database; 2) automated clustering of subsequent image cubes using the reference database, updating the database as needed to accommodate new spectral information; and 3) summary and creation of the final cluster maps. The algorithm considers both spectral shape and intensity, providing more accurate material classification. Detailed material identification for each cluster was then performed using complementary non-invasive spectroscopic techniques, including Raman spectroscopy, X-ray fluorescence (XRF), and high spectral resolution fiber optic reflectance spectroscopy (FORS), focusing on accessible areas. Principle Component Analysis (PCA) and Independent Component Analysis (ICA) were also applied to reveal 'hidden' writings and drawings not visible in individual spectral bands. The study compared the material composition of Cave 465 with those of dated Mogao caves from the Tibetan, Tangut, and Mongol/Yuan periods, using both data from the current study and previous research using X-ray diffraction (XRD) and XRF analysis of extracted samples. Palaeographic analysis of the revealed Sanskrit writings was also conducted to assist in dating.

The SOM clustering algorithm successfully grouped areas with similar spectra, even when widely separated, generating materials cluster maps. Over 10⁶ spectra from the east ceiling of Cave 465 were reduced to 960 clusters, with ~300 corresponding to physical damage. Multimodal, non-invasive analysis confirmed material identification from accessible areas, which were then extended to inaccessible areas based on the clustering results. The study identified a rich combination of pigments in Cave 465, including cinnabar, orpiment, indigo, azurite, gypsum, and red lead. The combination of azurite and indigo for dark blue and cinnabar and orpiment for orange appear characteristic of Tibetan paintings, while the indigo and orpiment combination for green suggests Nepalese influence. Comparison with other Mogao caves revealed that Cave 465 has a more complex pigment combination than other caves from the Tibetan, Tangut, and Mongol/Yuan periods. PCA and ICA analysis of spectral images revealed faded Sanskrit writings on the ceiling, identified as the "Summary of Dependent Origination." The script is consistent with a post-late 12th century date and is similar to those found in other caves dated to the Mongol/Yuan period. The archaeological evidence, along with the dated graffiti (1309-1373 AD), suggests a Mongol/Yuan date. Taking all evidence into account, including the material analysis and palaeographic analysis of the Sanskrit text, the study concludes that the main hall wall paintings of Cave 465 date to the late 12th to 13th centuries.

The findings demonstrate the power of combining advanced remote sensing, machine learning, and traditional methods in dating and understanding ancient artworks. The novel clustering algorithm allowed for efficient and accurate material identification on a large scale, overcoming limitations of previous methods. The interdisciplinary approach integrating material analysis, palaeographic analysis of the Sanskrit inscription, and archaeological evidence provided robust support for the dating of Cave 465 paintings. This study challenges previous dating arguments relying solely on stylistic comparisons or limited material analysis. The results highlight the potential for this approach in other cultural heritage projects, allowing for more comprehensive and precise analyses of large-scale artworks.

This study successfully dated the wall paintings of Mogao Cave 465 to the late 12th-13th centuries using a novel combination of remote sensing, machine learning, and traditional art historical and archaeological techniques. The developed SOM clustering algorithm proved highly effective in processing and analyzing large spectral imaging datasets, allowing for large-scale material identification. Future research could apply this method to other cave temples at Mogao and similar sites, further refining the understanding of the artistic styles and cultural exchanges along the Silk Road. Furthermore, developing more sophisticated machine learning models for spectral analysis and exploring the integration of other data sources (e.g., dendrochronology, radiocarbon dating where feasible) could enhance the accuracy and precision of dating ancient artworks.

While the study provides strong evidence for a late 12th-13th century date for Cave 465, some limitations exist. The dating of Tangut caves at Mogao remains uncertain, making definitive comparison challenging. The analysis relied on non-invasive techniques, limiting the ability to analyze the full range of pigments and dyes, especially organic ones. While the study used a large dataset, it still represents only a portion of the Mogao caves. Finally, the accuracy of the dating relies on the accuracy of the dating of comparison caves and the existing scholarship regarding their attributes.